[Juliusolsson05]

Closes #1113. Carries the package fix for #1114 (Juliusolsson05/opencode-terminal-headless#6).

Both bugs were diagnosed from the debug bundle the owner captured at the moment they happened (2026-09-21T22-09-23-834-168b7a32, note: "Sending logic broke") plus the app's own incident log.

---

1. Claude: a word-wrapped [Image #N] pill makes image prompts fail as "never submitted"

22:06:59.341  submit.begin   {provider: claude, source: with-images}
22:07:00.360  input.write    {origin: delivery, writes: 3, bytes: 178}
22:07:04.369  submit.result  {ok:false, code:"absorption-timeout", stage:"absorption",
                              bodyWritten:true, enterWritten:false, retryable:false, durationMs:5027}

The prompt and the image were in the composer. The recorded screen at that instant:

❯ Lets figure out why this is happening wiht opencode [Image
  #1]

imagePlaceholderCount matched /\[Image #\d+\]/ against the raw composer. The pill has exactly one internal space, and a word wrap can only break a line at a space — so when the pill lands on the wrap column, neither half matches. The count never rises above its baseline, pollClaudeImagesAbsorbed spends its whole 5 s budget, and delivery returns absorption-timeout with retrySafe: false. The user gets "the prompt was never submitted — resend is blocked" and has to clear the agent's composer by hand.

Counted across every frame in the bundle:

frames	literal match	whitespace-normalized
seq 21–23 (the failing send)	0	1
seq 51–59 (the send that worked)	1–2	1–2

Same pane, same session, same terminal width, ten minutes apart. The only difference is where that one space fell — which is why this presented as intermittent rather than as a broken feature.

pasteAbsorbedVia, the text half of the same protocol, has always compared against normalizeWhitespace(screen). That is precisely why wrapped text sends never had this problem. The two halves had drifted; that inconsistency was the defect. Both placeholder counters now normalize.

placeholderCount is fixed by analogy, not from a recording — and its failure mode is worse: [Pasted text #1] has two spaces to wrap at, and a collapsed paste has no inline tail to fall back on, so a missed placeholder there is a guaranteed timeout with the draft stranded.

Tests

testing/fixtures/image-absorption/ holds two real baseline/after pairs from the bundle — one wrapped delivery that failed, one unwrapped delivery that succeeded — and promptDelivery.imageWrap.recorded.test.ts replays them through deliverClaudePrompt, the real main-process entry point, stubbing only the PTY sink and the JSONL acceptance observer.

This is a recorded test rather than a screen literal on purpose: anyone writing a fake screen puts [Image #1] on one line, which is exactly the case that always worked. A literal would have passed against the broken detector.

Fail-first was verified: before the fix the wrapped case failed at 5025 ms with absorption-timeout while unwrapped passed — so the test isolates the wrap and nothing else.

---

2. OpenCode: a 20 s db-path timeout disabled the durable channel permanently

OpenCode durable channel (db_path_unavailable):
`…/runtime/opencode/darwin-arm64/opencode db path` failed: Command failed: …/opencode db path

The binary is fine — I measured it at 0.37 s warm, 2.15 s cold. The app's incident log shows the timeout firing:

when	provider.start.end (opencode)
2026-09-20T22:52:30.453Z	20087 ms (= the 20 000 ms budget + overhead)
2026-09-21T17:14:03.303Z	24095 ms
2026-09-21T17:14:03.304Z	23830 ms

Healthy runs in the same log: 4–2300 ms.

Full analysis and the fix are in the package PR. In short: a ~143 MB Bun binary started several times over during a multi-pane restore overruns a flat 20 s budget; execFile's kill yields Command failed: <cmd> with empty stderr (hence the unactionable banner); the resolver memoises so every pane in the window shares the one rejection (both panes above report in the same millisecond); and openDurable() treated a null path as terminal, so one slow startup cost those panes their committed transcript until the app restarted.

The package now distinguishes a timeout from a bad install from a spawn failure in the message, and retries the lookup in the background on its own ladder. This repo's side is one thing: passing the resolver, since the app owns every child process — the package never spawns one.

---

Verification

• npx tsc -b clean on both projects.
• 320/320 in opencode-terminal-headless; 5525 passed in the app suite.
• One unrelated failure, src/main/workflows/control.system.test.ts, reproduced on origin/main in a clean worktree — it is the wall-clock flake tracked in bug(test): two system tests synchronize on wall-clock time and fail under load #1107, not this change.

🤖 Generated with Claude Code

[Juliusolsson05]

Adversarial review round — two reviewers

No behavioural defect found in the fix. Both reviewers verified the fail-first claim by reverting the one-line change and watching the wrapped case die at ~5 s while unwrapped stayed green, and both verified the fixture against its source bundle byte-for-byte (frames, screenHash, tsIso, and both submitResult blocks against incident-events.jsonl). One went further and proved imageCount: 1 from the recorded input.write byte counts rather than taking it on trust.

One reviewer swept 10,263 recorded frames from 72 debug bundles through the old and new counters. Exactly 4 diverge — all four are the bug frames from the source bundle. Zero paste-counter divergences across 78 frames containing [Pasted. They also measured the composer bound rather than assuming it: of 5,835 frames containing ❯, 5,827 resolve through the divider-bounded path, so scrollback leakage is not a live path.

The change also closes a pre-existing hole: a wrapped pill previously counted 0 at baseline, so a later unwrapped pill could confirm on a 0→1 delta that was not a transition at all.

What did need fixing: three comments that state falsehoods

In a repo whose CLAUDE.md makes the comment the deliverable, these are findings, not nits.

• "pasteAbsorbedVia has always normalized, which is why wrapped TEXT sends never carried this bug." Wrong, and it contradicted a comment 25 lines above it. Only the inline branch normalized; the placeholder branch matched the raw screen exactly as the image counter did. Inlined pastes were immune; collapsed ones were not — and a collapsed paste has no inline tail to fall back on, so a missed placeholder there is a guaranteed timeout. The sibling change is a fix, not a precaution.
• "A word wrap can only break at a space." Stated without its bound. Claude's Ink calls wrapAnsi(…, { hard: true }) (vendor/claude-code-src/full/ink/wrap-text.ts), so a token longer than the pane is broken mid-token. For a 10-character pill that needs a composer under ~11 columns, which is why the premise holds here — but it does not hold for the inline tail needle. Now bounded explicitly.
• "\s+ catches a pill that survives normalization with an odd gap." Nothing does: after normalizeWhitespace no run longer than one space can exist, so \s+ is provably equivalent to a literal space here.
• The recorded test's "replaying after for both phases would quietly test nothing" — it FAILS loudly instead, which the reviewer confirmed by doing it. The reconstruction is still required, for a different reason.

A real bug the review surfaced, filed rather than fixed here

#1118 — the inline tail needle is defeated by a hard wrap inside a long token. Driving the real detector over a width sweep, a 110-char prompt ending in an absolute path never absorbs at pane widths 46–50 and 90–98 (14 of 81). Byte-for-byte the #1113 experience, on the text path. Different signal, different fix.

#1119 — deliverClaudeImagePrompt takes the image baseline before the short-text write has painted, so a prompt whose own text contains [Image #N] confirms before any attachment exists. Pre-existing; this PR widens it only in that a wrapped literal now counts. The docstring's delta-safety argument answered the wrong axis and now says so.

Fixture README

Claimed "nothing else differs" between the two deliveries. Disproved point by point: different routes (51 vs 120 chars → raw write vs bracketed paste), idle vs mid-turn, and the pane was resized 64 → 120 columns between them. The conclusion survives — both got their text into the composer, so image absorption alone decided their fate — and the resize frame actually strengthens the story. Also corrected: 2 m 17 s apart, not ten minutes; three image deliveries in the bundle, not two.

One more thing the review caught

Both Claude paste debug probes still defaulted to the raw literal, so the first tool anyone opens on the next stranded-draft report would have said "0 placeholders" for exactly the screen production now reads as 1. Fixed.

tsc -b clean; 5527 app tests pass. The one failure, control.system.test.ts, reproduces on a clean origin/main worktree (#1107).

🤖 Generated with Claude Code